CN101570788A - Method for recognizing genotype through single nucleotide polymorphism chip - Google Patents

Abstract

The invention provides a method for recognizing genotype through a single nucleotide polymorphism (SNP) chip. The method comprises the following steps: 1, performing arrangement and dimensionality reduction treatment on optical density signals in a genetic chip, and reducing the dimensionality of metering data of each SNP from 20 dimensionalities to 2 dimensionalities; 2, performing guiding-free classification, primarily judging the genotype by Mahalanobis distance, and recognizing the genotype of partial SNP according to the size of the Mahalanobis distance between two-dimensional data and zeropoint and the positive/negative condition of a median of a difference value; and 3, performing guiding classification, and performing further judgment by utilizing the information of a training sample if the genotype cannot be determined in step 2. The method has the advantages that: in the dimensionality reduction process, the influence of a probe position factor is eliminated, the degree of variation of the data is reduced, the accuracy is improved, the primary guiding-free classification judgment determines the genotype of over 50 percent of the SNP, defects of other methods that the CRLMM and the like have too large calculation amount and are difficultly realized on a common computer are overcome, and high accuracy is ensured under the condition of small samples.

Description

A kind ofly discern genotypic method by single nucleotide polymorphism chip

Technical field

The present invention relates to genotypic identification, particularly a kind ofly discern genotypic method by single nucleotide polymorphism chip, this method is applicable to the genotype identification of small sample and flow process and the time that can greatly simplify the identification of large sample genotype simultaneously.

Background technology

Single nucleotide polymorphism (SNP) gene chip is clocklike to be arranged on the substrate with the SNP oligonucleotide chain (to call probe in the following text) of knowing dna sequence dna, will hybridize by the base complementrity pairing with probe behind the testing sample dna marker.Scan hybridizing the back chip by fluorescence detecting system, the optical signal data that obtains is handled and analyzed, just can identify the SNP genotype of sample.Utilize gene chip, can be simultaneously 100,000 even above different SNP be carried out genotypic identification and detection, thereby be the downstream gene engineering, as the detection of disease gene, gene therapy etc. provide strong data support.The SNP chip is widely used in life science and practice, fields such as medical research and clinical medicine design now as high-throughput genotype identification facility now.Affymetrix SNP chip comprises GeneChip Human Mapping 100k microarray and 500k microarray, especially because their stdn and automatization become one of popular research platform the most.Along with the continuous increase of microarray density, this has brought new challenge also for the data processing and the statistical inference of chip.

In SNP genotype recognition process, a most basic problem is how to provide an accurately effective method of discrimination.For early stage Affymetrix 10k SNP chip, can use and revise stripping (Modified Partitioning Around Medoids, MDAM) (Liu, W., Di, X., Yang, G., Matsuzaki, H., Huang, J. (2003) Algorithms for large-scale genotypingmicroarrays, Bioinformatics, 19,2397-2403.), but this method is under the lower situation of some allelotrope frequency of occurrences, and the result can be accurate inadequately, and this problem seems even more serious under 100k and a large amount of situations about using of 500k chip.2005, Affymetrix company proposed a kind ofly to be called dynamicmodel (this model is genotype recognition result and the corresponding reliability level (Di that obtains by the Wilcoxon signed rank test for Dynamic Model, DM) method, X., Matsuzaki, H., Webster, T.A., Hubbell, E., Liu, G., Dong, S., Bartell, D., Huang, J., Chiles, R., Yang, G., Shen, M.M., Kulp, D., Kennedy, G.C., Mei, R., Jones, K.W., Cawley, S. (2005) Dynamic model based algorithmsfor screening and genotyping over 100K SNPs onoligonucleotidemicroarrays, Bioinformatics, 21,1958-1963.).Though DM does not need learning sample, the heterogeneous genotype of tending to fail to judge.Popular along with 100k and 500k chip, a lot of novel methods occur one after another.Rabbee and Speed have proposed sane linear model method (the Robust Linear Model with Mahalanobis Distance Classifier based on the mahalanobis distance classification, RLMM), it uses the HapMap data as learning sample (Rabbee, N., Speed, T.P. (2006) Agenotype calling algorithm for affymetrix SNP arrays, Bioinformatics, 22,7-12.).Affymetrix company improves at RLMM, developed Bayes's robust linear regression method (Bayian Robust Linear Model with MahalanobisDistance Classifier based on the mahalanobis distance cluster, BRLMM), this method uses the result of DM algorithm to eliminate the difference of different experiments as learning sample.Carvalho etc. have proposed the sane linear model method of correction (the Corrected Robust Linear Model with Maximum Likel ihoodClassification based on the maximum likelihood classification, CRLMM), obtained all well and good result (Carvalho, B., Speed, T.P., Irizarry, R.A. (2007) Exploration, Normalization, and Genotype Callsof High Density Oligonucleotide SNP Array Data, Biostatistics, 8,485-499.).But because CRLMM need carry out pre-treatment to used chip data, therefore it requires computer to have very high service ability and handles the lot of data operation, and when new sample occurs, again carrying out pre-treatment may obtain and different genotypic differentiation result last time, more seriously, the CRLMM method also loses tolerance range easily under small sampling condition.

Summary of the invention

The purpose of this invention is to provide and a kind ofly discern genotypic method by single nucleotide polymorphism chip, this method is simple to operation and kept very high tolerance range, do not need that data are carried out regularization and handle, and to a certain extent need be by other sample informations.

The object of the present invention is achieved like this:

Step 1, original signal arrangement dimensionality reduction, optical density signal in the gene chip is carried out dimension-reduction treatment, be divided into 4 groups according to the different signal densities of same SNP of will representing of allelic difference with the positive minus strand of probe bonded DNA, subtract each other by the correspondence between the isoallele not and to become two groups, become the two-dimensional signal difference again, get the median of two groups of genotype signal logarithm differences again;

Step 2 does not have the directiveness classification, according to mahalanobis distance principium identification genotype, according to the signal difference median and zero point mahalanobis distance size and the positive and negative situation identification division SNP genotype of signal difference median;

Step 3 has the directiveness classification, if the principium identification algorithm be can not determine genotype, then utilizes the information of learning sample further to differentiate, and utilizes the center of the mean vector of learning sample different genotype as classification, sorts out differentiation.This step comprises the calculating of different genotype difference in signal strength average, prior probability and revises the calculating of mahalanobis distance; Judge genotype according to minimum correction mahalanobis distance criterion.

It is that the mean value computation of the SNP optical signal logarithmic value that equated with the minus strand number of probes by normal chain probe in the chip obtains that described mahalanobis distance calculates the variance of using.

Described different genotype difference in signal strength average is to obtain by the average of adding up existing cdna sample.

Described prior probability is this genotype SNP proportion in known type training set.

Described correction mahalanobis distance is that the prior probability natural logarithm value that general mahalanobis distance deducts twice is obtained.

The invention has the advantages that: eliminated the influence of probe location factor in the dimensionality reduction process, reduced the degree of variation of data, improved accuracy rate, preliminary nothing directiveness is classified and has been determined the genotype of 50% above SNP with 99.95% high confidence level, has saved the discriminant classification time greatly.

Description of drawings

Fig. 1 is the optical density signal logarithmic value broken line graph of the not homoallelic same position probe of SNP

Fig. 2 is the logarithmic signal difference median scatter diagram of all SNP in a certain chip

Fig. 3 calculates the ratio that abandons and the accuracy rate broken line graph of 3 samples for the present invention and CRLMM method

Fig. 4 is a schema of the present invention

Embodiment

In conjunction with the accompanying drawings embodiments of the invention are elaborated, present embodiment is implemented under the prerequisite with technical solution of the present invention, provided detailed embodiment and concrete operating process, but protection scope of the present invention is not limited to following embodiment.

I), definition

The data of gene chip obtain by fluorescing system scanning, and the file of storage is that suffix is the CEL form.Remove some descriptive data such as gene chip model, SNP title in the chip, sample title etc., the signal density data are stored in the file with the form of matrix.Two kinds not isoallele remember into A and B respectively, the two strands of DNA is remembered respectively and is done normal chain and minus strand.Basically each SNP has four types data according to probe kind difference, is respectively the probe data that detects A type SNP normal chain, detects the probe data of Type B SNP normal chain, detects the probe data of A type SNP minus strand and the probe data of detection Type B SNP minus strand.Every type probe wherein, according to target nucleotide sequence position different designs a plurality of probes, in order to eliminate experimental error, the base coupling probe (perfect matach probe) of quantity and the base probe (mismatch probe) that do not match such as designed simultaneously.This law is only considered base coupling probe.Number of probes is by shown in the table one, and in dissimilar chips, quantity does not wait; And in chip of the same race, the number of probes of different SNP and dissimilar probe ratio may not wait yet.Remember y respectively _{A+, j}And y _{A-, j}Original signal density for the j bar probe of allelotrope A normal chain and minus strand; Accordingly, for allelotrope B, note is y respectively _{B+, j}And y _{B-, j}The y value is got 2 to be the logarithm at the end, and we obtain as drag

$\left\{\begin{array}{c}{\log }_2\left(y_{A+,j}\right)={\mathrm{\θ}}_{A+}+T_{+,j}+{\mathrm{\ϵ}}_{A+,j},j=1,...,s_{A+,}\\ {\log }_2\left(y_{A-,j}\right)={\mathrm{\θ}}_{A-}+T_{-,j}+{\mathrm{\ϵ}}_{A-,j},j=1,...,s_{A-,}\\ {\log }_2\left(y_{B+,j}\right)={\mathrm{\θ}}_{B+}+T_{+,j}+{\mathrm{\ϵ}}_{B+,j},j=1,...,s_{B+,}\\ {\log }_2\left(y_{B-,j}\right)={\mathrm{\θ}}_{B-}+T_{-,j}+{\mathrm{\ϵ}}_{B-,j},j=1,...,s_{B-,}\end{array}\right.$

Wherein θ has represented a certain allelotrope (A or B) and a certain chain (normal chain or minus strand) information strength down, T represent the target nucleotides sequence be listed in difference generation in present position in the probe position effect, ε represents measuring error.Because for every kind of allelotrope, number of probes equates, as s _A+=s _B+, s _A-=s _B-So,, we will simply remember into s ₊, s _-

Ii), data preparation and dimensionality reduction

At first, we wish that generally normal chain equates with the number of probes of minus strand, i.e. s ₊=s _-But in fact, in chip design and the making processes, because some restriction, often can not realize that there are the probe of as much in normal chain and minus strand, small portion SNP in SNP in the 100k chip and the 500k chip, its normal chain and minus strand probe ratio did not wait from 3: 7 to 7: 3, the most of SNP in the 500k chip, its normal chain and minus strand probe ratio did not wait from 1: 5 to 6: 0, and be as shown in table 1.And positive minus strand equilibrated probe groups (ratio 5: 5 or 3: 3), its quantity is less than half of total probe groups.The second, because same group of probe have the different probe of a plurality of target nucleotide positions, and on the same group the probe signals of probe same target nucleotide position does not have metastable signal difference, as shown in Figure 1.Not homoallelic same position probe signals presents the parallel construction feature, and based on such phenomenon, we define

$\left\{\begin{array}{c}{M}_{+,j}={\log }_2\left(y_{A+,j}\right)-{\log }_2\left(y_{B+,j}\right),j=1,...,s_{+},\\ M_{-,j}={\log }_2\left(y_{A-,j}\right)-{\log }_2\left(y_{B-,j}\right),j=1,...,s_{-},\end{array}\right.$

This model has at first been eliminated the position influence of probe signals, and then passes through { M _{+, j}, j=1 ..., s ₊And { M _{-, j}, j=1 ..., s _-Get median respectively and obtain M ₊And M _-, as real signal difference θ _A+-θ _B+And θ _A--θ _B-Estimation.For the probe groups that lacks normal chain or minus strand (as the 6:0 type), we can use M ₊Estimate M _-

Suppose M={M ₊, M _-The obedience average is μ={ μ ₊, μ _-, variance is the two-dimentional normal distribution of ∑.Obviously, M ₊With M _-Positive correlation, { the μ of frequency of genotypes AA _{AA ,+}, μ _{AA ,-}Be the positive number, { μ of genotype BB _{BB ,+}, μ _{BB ,-}Be negative, and { the μ of genotype AB _{AB ,+}, μ _{AB ,-}Then be near the initial point.{ M with all SNP in the chip ₊, M _-Do scatter diagram (as Fig. 2) discovery, average obviously is divided into foregoing three groups, and is positioned on the diagonal lines.We suppose simultaneously, and variance matrix and position effect are independent, depend on the influence of different experiments.We use the mean value of the variance of all equilibrated probe groups to make estimation to variance, and equilibrated probe groups variance can be by paired { M _{+, j}, j=1 ..., s ₊And { M _{-, j}, j=1 ..., s _-(this moment s ₊=s _-) estimate.Estimation as shown in table 1, as to have 36% SNP can be used for calculating ∑ in the 100k probe, and in the 500k probe, can calculate ∑ by weighted mean simultaneously in conjunction with two kinds of balance probe groups (3: 3,5: 5).

The number of probes of table 1 normal chain and minus strand is than (s ₊: s _-)

(measuring an allelic situation in the 500k chip by 6 PM probes in the bracket)

Iii), calculate

Definition mean vector M={M ₊, M _-And actual value between mahalanobis distance be:

d(M，μ)＝(M-μ)∑ ^-1(M-μ)

Under normality assumption, (M, μ) the obedience degree of freedom is card side's distribution of 2 to d, remembers χ respectively _{2, α} ²And χ ₂ ²(x) be α percentage point and the distribution function that 2 card side distributes for degree of freedom.If 1-α is a significance level.The two stage genotype method of identifications that we propose are as follows:

(1), the fs, use guideless classification method

If a $d(M,0)<{\mathrm{\&chi;}}_{2,\mathrm{\&alpha;}}^2,$ Then declaring SNP is the AB type, and reliability level is 1-χ ₂ ²(d (M, 0));

If b $d(M,0)>{\mathrm{\&chi;}}_{\mathrm{2,1}-\mathrm{\&alpha;}}^2,$ And (M ₊, M _-) just be, then declaring SNP is the AA type, reliability level is 1-χ ₂ ²(d (M, 0));

If c $d(M,0)>{\mathrm{\&chi;}}_{\mathrm{2,1}-\mathrm{\&alpha;}}^2,$ And (M ₊, M _-) be negative, then declaring SNP is the BB type, reliability level is 1-χ ₂ ²(d (M, 0)).

(2) otherwise, enter subordinate phase, the instruction classification method is utilized learning sample information

A, use HapMap data obtain the mean vector of genotype g as learning sample, are defined as

μ _gThe average of the M value of=known learning sample

And the priori gene frequency of corresponding genotype g is p _g, be defined as

g＝AA，AB，BB

B, calculate each genotypic correction mahalanobis distance:

$\tilde{d}(M,{\mathrm{\&mu;}}_g)=d(M,{\mathrm{\&mu;}}_g)-2\ln \left(p_g\right)$

C, differentiation genotype are $g^{*}=\arg \min \tilde{d}(M,{\mathrm{\&mu;}}_g),$ Reliability level is 1-χ ₂ ²(d (M, μ _g)).

Some precaution: 1. the classification of fs identification is simply fast.2. in subordinate phase, the allelotrope frequency of occurrences of learning sample is considered and studies positive mahalanobis distance, is equivalent to Bayes's posterior probability.3. for having lacked μ in some learning sample _AAOr μ _BBSituation, we are with the known μ of allelic other all SNP of all same types _gAverage estimate, as μ _{G, missing}=all μ that observe _gAverage, and for μ _AB, we simply use, and (O O) replaces.When new sample genotype information, learning sample can upgrade accordingly.4. reliability level can be used to screen suitable SNP, carries out the research of the genetics in downstream.If 5. learning sample does not exist, the result who is obtained by fs identification can be used as new learning sample, for subordinate phase is sorted out service.On this degree, our method is self to adapt to and independently.

Iv, application examples

The genotype identification of 30 samples of Europe ethnic group

In this example, 30 chips are to be the part sample of enzyme with Hind among the Gene Chip Human Mapping 100k.The learning sample that uses is the information of 269 samples of HapMap 100K data.The several SNP that below select one of them sample describe recognition process in detail:

Sample title NA * * * * * * * * *,

Step 1, reading of data, getting the signal value y of chip results with 2 is the logarithm log at the end ₂(y), each SNP is divided into 4 groups according to the different signal densities of same SNP of will representing with the positive minus strand of probe bonded DNA of allelic difference

Step 2, data are put in order and dimensionality reduction in earlier stage, calculate M respectively _{+, j}=log ₂(y _{A+, j})-log ₂(y _{B+, j}), j=1 ..., s ₊And M _{-, j}=log ₂(y _{A-, j})-log ₂(y _{B-, j}), j=1 ..., s _-Median (M ₊, M _-)

The SNP name	{M +，j，j＝1，...，s +}	{M -，j，j＝1，...，s -}	M +	M -
					SNP_A-1712762	2.18 2.04 1.92 2.61 2.01 2.09	2.58 2.94 3.20 3.06	2.06	3.00
SNP_A-1718890	-0.61 -0.71 -1.13 -0.42 -0.68	-0.69 -0.28 0.13 0.04 -0.70	-0.68	-0.28
					SNP_A-1668776	-2.00 -1.95 -0.64 -1.62 -1.95	-0.84 -1.06 -0.40 -1.34 -0.66	-1.95	-0.84
SNP_A-1723597	2.09 2.49 2.93 2.87 2.39 2.43 1.92	1.90 2.18 2.21	2.43	2.18
					SNP_A-1728870	3.21 3.12 2.68 3.26 3.21	1.61 2.44 1.40 1.29 1.54	3.21	1.54

Step 3 is to each chip, according to (the M of balance probe groups _{+, j}, M _{-, j}), j=1 ..., s, s=s ₊=s _-, calculate and estimate its covariance ∑ of SNP i _i, the computing chip population variance is estimated again, $\widehat{\mathrm{\&Sigma;}}={\mathrm{\&Sigma;}}_i$ Average.

Step 4 is calculated $d(M,0)=(M-0){\widehat{\mathrm{\&Sigma;}}}^{-1}{(M-0)}^{\&prime;},$ Differentiate itself and χ _{2, α} ²And χ _2,1-α ²Relation, make α=0.0005, ${\mathrm{\&chi;}}_{2,\mathrm{\&alpha;}}^2=0.00100025,$ ${\mathrm{\&chi;}}_{\mathrm{2,1}-\mathrm{\&alpha;}}^2=15.2018.$ Carry out principium identification.The horizontal 1-χ of while computed reliability ₂ ²(d (M, 0))

The SNP name	d(M，0)	With χ 2，α 2And χ 2，1-α 2Relation	Differentiate genotype
				SNP_A-1712762	56.19060	Greater than χ 2，1-α 2	AA
SNP_A-1718890	2.395623	Between	Undetermined
				SNP_A-1668776	20.03103	Greater than χ 2，1-α 2	BB
SNP_A-1723597	45.61602	Greater than χ 2，1-α 2	AA
				SNP_A-1728870	56.09637	Greater than χ 2，1-α 2	AA

Step 5 is calculated $\tilde{d}(M,{\mathrm{\&mu;}}_g)=d(M,{\mathrm{\&mu;}}_g)-2\ln \left(p_g\right),$ Differentiating genotype is $g^{*}=\arg \min \tilde{d}(M,{\mathrm{\&mu;}}_g),$ Reliability level is 1-χ ₂ ²(d (M, μ _g)).

Gained is listed in tolerance range such as the table 2 as a result, provides simultaneously and additive method result relatively.As can be seen, the tolerance range of CRLMM has reduction largely under the little situation of sample size.And the present invention is insensitive to the sample size size when keeping accuracy.As can be seen, under 3 sample situations, when the ratio that abandons is identical (ratio of abandoning refers to abandon those minimum SNP of a certain proportion of reliability level), accuracy of the present invention is better than CRLMM all the time from accompanying drawing 3.

Table 2

Sample size	Chip type	The present invention	CRLMM
					3	Hind	99.86％	99.23％
	Xba	99.69％	99.20％
				6	Hind	99.87％	99.68％
	Xba	99.71％	99.65％
				12	Hind	99.80％	99.82％
	Xba	99.72％	99.82％
				30	Hind	99.81％	99.87％
	Xba	99.77％	99.87％

Claims (5)

1, a kind of by the genotypic method of single nucleotide polymorphism (SNP) chip identification, it is characterized in that this method may further comprise the steps:

A), the optical density signal in the gene chip that defines model is put in order dimension-reduction treatment

The signal density of the same SNP of representative is divided into 4 groups according to allelotrope sequence position different and different with the positive minus strand of probe bonded DNA on probe, subtract each other by the respective signal density between the isoallele not and to become two groups, become the two-dimensional signal difference again; The two-dimensional signal difference is the median of two groups of genotype signal logarithm differences;

B), there is not the directiveness classification, with mahalanobis distance principium identification genotype

With mahalanobis distance and threshold ratio, the positive and negative situation of judgment signal difference median; Its mahalanobis distance is counted to the mahalanobis distance of initial point for the signal difference meta;

C), directiveness classification is arranged, if principium identification be can not determine genotype, then utilize the information of learning sample further to differentiate

Comprise the calculating of different genotype difference in signal strength average, prior probability and the calculating of correction mahalanobis distance; Judge genotype according to minimum correction mahalanobis distance criterion.

2, the method for claim 1 is characterized in that it is that the mean value computation of the SNP optical signal logarithmic value that equated with the minus strand number of probes by normal chain probe in the chip obtains that described mahalanobis distance calculates the variance of using.

3, the method for claim 1 is characterized in that described different genotype difference in signal strength average is to obtain by the average of adding up existing cdna sample.

4, the method for claim 1 is characterized in that prior probability is this genotype SNP proportion in known type training set.

5, the method for claim 1 is characterized in that described correction mahalanobis distance is that the prior probability natural logarithm value that general mahalanobis distance deducts twice is obtained.

Images